H100 Price Factors: What Moves a Reserved Capacity Quote

Picture a procurement lead at a mid-sized AI startup. Engineering has asked for H100 capacity for a fine-tuning program, finance wants a number by Friday, and two quotes have landed in her inbox. They look similar at a glance, but one assumes a longer term, one bundles a different interconnect, and neither says clearly who installs the drivers. She cannot tell which is cheaper because she cannot tell whether they describe the same thing.

This guide is for buyers in that seat: procurement, infrastructure and finance leads evaluating a reserved H100 quote. It will not give you rates. It explains the factors that move an H100 price on a reserved commitment, so you can read a quote, ask the right questions and compare offers on equal terms. For broader context on how dedicated bare-metal rental works across on-demand and reserved terms, start with Pacific's guide to dedicated GPU rental and reserved terms.

1. Start with what the quote actually covers

Before you compare numbers, confirm the unit. A reserved H100 quote can be expressed per GPU, per node or per cluster, over a term, with or without networking, storage and support folded in. Two quotes that look far apart may be pricing different bundles.

General cloud cost drivers like egress, storage tiers and utilization patterns are covered in Pacific's guide to GPU cloud pricing cost drivers. This post narrows the focus to the specific levers inside a reserved H100 quote.

• Is the quote per GPU, per node or for the full allocation?
• What term does it assume, and when does that term start?
• Which node configuration, memory and interconnect are included?
• What does the provider own operationally, and what falls to your team?
• Is the capacity confirmed, or is it an indicative figure pending review?

2. Term length is the first lever

Term length shapes a reserved quote more than almost anything else. A provider is setting aside specific hardware for you, and the length of that commitment affects how they plan the rest of their fleet. Longer commitments give the provider more certainty, and shorter ones give you more flexibility. The quote reflects that trade.

On the Pacific Compute catalog of dedicated bare-metal GPU nodes, the H100 listing offers on-demand access plus reserved terms of 7-30, 31-90 and 91-180 days by request. Those bands are useful because they force a planning question: how long does your workload really need the hardware?

• Map the term to a real milestone, such as a training run, an evaluation cycle or a product launch window.
• Ask what happens at the end of the term: extension options, notice periods and whether the same nodes stay with you.
• Check whether the term clock starts at order acceptance, at handover or at first access.
• Avoid buying a longer band only because it looks efficient on paper. Unused reserved time is still spent budget.

3. Node configuration sets the baseline

The second factor is the hardware itself. "H100" covers more than one form factor and memory configuration, and providers package GPUs into nodes in different ways. A quote that does not name the exact configuration leaves room for surprises.

Pacific's H100 listing is specific: H100 SXM with 80 GB of memory, 8 GPUs per node, delivered as dedicated bare-metal Linux nodes with SSH access. That level of detail is what you want in any quote you evaluate, because it lets you compare like with like.

• Confirm the form factor (SXM or another variant) and the memory per GPU.
• Confirm GPUs per node and the number of nodes in your allocation.
• Ask whether nodes are dedicated bare metal or shared and virtualized.
• Check local storage: type, capacity and what happens to data when the resource ends. On Pacific's H100 listing, local NVMe is erased when the resource ends.
• Make sure the node count matches how your jobs actually scale, since buying partial nodes is rarely an option.

If you are still deciding whether H100 is the right generation for the workload, the comparison in B200 vs H100 for reserved capacity walks through how the choice changes the reservation conversation.

4. Interconnect changes what the nodes can do together

For single-node work, interconnect barely matters. For distributed training across many nodes, it determines whether your GPUs can work as one system or spend their time waiting on each other. That is why interconnect is a pricing factor: a high-bandwidth fabric adds cost and planning to the provider's side, and it changes what your workload can accomplish.

Pacific's H100 listing specifies InfiniBand interconnect. When you compare quotes, confirm the fabric explicitly instead of assuming it. A quote with a lower headline figure may describe nodes connected only by standard Ethernet, which can be fine for inference or independent jobs and a poor fit for tightly coupled training.

• Name the interconnect in writing: InfiniBand, Ethernet or another fabric.
• Ask how nodes in your allocation are connected to each other, and whether the fabric is shared with other tenants.
• Match the fabric to the workload. Multi-node training needs it, while batch inference often does not.
• Check whether storage traffic shares the same network as GPU-to-GPU traffic.

5. The support boundary is part of the price

Every quote implies a line between what the provider operates and what you operate. That line has real cost on both sides. If the provider owns more, it shows up in the quote. If your team owns more, it shows up in your engineering hours.

On Pacific's H100 listing, nodes arrive with NVIDIA drivers and CUDA installed. Your team brings the stack above that: frameworks, containers, orchestration, job scheduling and your own monitoring. Knowing that boundary ahead of time lets you staff for it and avoid paying twice for the same work.

• List what is preinstalled: operating system, drivers, CUDA version.
• Ask who handles hardware faults, node replacement and firmware updates.
• Clarify who owns the software stack, scheduler and cluster tooling.
• Confirm how support requests are raised and what response expectations are written into the terms.
• Estimate the internal effort your team will spend on everything above the provider's line, and include it in your comparison.

6. Delivery window confirmation decides whether the quote is real

The most overlooked pricing factor is the status of the quote itself. A figure attached to unconfirmed capacity is an estimate. A figure attached to an accepted order with a confirmed delivery window is a commitment. Treat them differently in your budget.

Pacific is explicit about this. Availability shown in the catalog is recorded by Pacific and confirmed when an order is accepted. Preview rates are indicative until Pacific confirms the allocation and are not a live inventory feed. Before you commit, review revalidates capacity, the current quote and terms. And a paid order on its own does not mean the resource is ready for your team to use.

• Ask whether the quoted figure is indicative or confirmed.
• Find out what the review step checks and what could change between quote and acceptance.
• Get the delivery window in writing at order acceptance.
• Separate payment from readiness in your internal plan. Schedule workloads against the confirmed handover date rather than treating payment as readiness.
• Keep a short log of what was confirmed, by whom and when, so finance and engineering share one source of truth.

7. Build a like-for-like comparison sheet

Once you understand the factors, put every quote into the same frame. A simple comparison sheet keeps you from rewarding the quote that left the most out.

• Unit: per GPU, per node or per allocation.
• Term band and start trigger.
• Hardware: form factor, memory per GPU, GPUs per node, node count.
• Interconnect and whether it is dedicated to you.
• Local storage and data handling at end of term.
• Support boundary: what is preinstalled, what the provider maintains, what you own.
• Quote status: indicative or confirmed, and what review revalidates.
• Delivery window and how it is confirmed.
• Internal cost of the work your team takes on.

If your roadmap includes newer hardware later, it can help to see how the same factors play out on a different generation. H100 vs B300 for reserved capacity covers how the reservation process differs when you move from catalog nodes to a larger by-request build.

If you want help mapping your workload to a term band, node count and interconnect before you request a reserved H100 quote, book a 30-minute call with the team at Pacific Intelligent Technologies, Inc.. Bring your workload profile and target milestones, and you will leave with a clear list of what to confirm in writing.

FAQ

Why won't this guide tell me the H100 price?

Reserved pricing depends on term, configuration, interconnect, support scope and confirmed availability at the time of review. A single figure without those inputs would mislead more than it helps. Understanding the factors lets you read any quote accurately and push back where a quote is vague.

What reserved terms are available for H100 on Pacific Compute?

The H100 listing offers on-demand access plus reserved terms of 7-30, 31-90 and 91-180 days by request. Choose the band that matches a real workload milestone, then confirm start conditions and end-of-term options during review.

Does paying for an order mean my H100 nodes are ready?

No. Availability is confirmed when an order is accepted, and review revalidates capacity, the current quote and terms before you commit. Payment is one step in the process. Plan your workload start around the confirmed delivery window and handover, and keep those dates separate from the payment date in your internal schedule.